Neural Networks

What are Neural Networks?

A neural network is a function approximator built from layers of simple units. Each unit computes a weighted sum of its inputs, adds a bias, and passes the result through a nonlinear activation function; stacking such units in layers lets the network represent nonlinear relationships that a single linear model cannot. The weights are learned by minimizing prediction error on historical examples, almost always by gradient descent with backpropagation supplying the gradients. In trading tools the usual form is a small multilayer perceptron (MLP): an input layer of hand-built features, one or two hidden layers, and an output read as a forecast value or, through a sigmoid or softmax, as class probabilities.

Flexibility is both the appeal and the hazard. A network with enough parameters can fit training data almost perfectly by memorizing noise, so in-sample accuracy means little; what counts is performance on data the model never saw, which makes an in-sample / out-of-sample split or walk-forward test non-negotiable. Chart-based implementations add constraints of their own: small datasets, features limited to what the platform computes, and retraining schemes that can quietly leak future information. A neural network is a fitting method, not a strategy; its value is set by the features it is fed, the label it is trained to predict, and the honesty of its validation.

Architecture families matter less on charts than in research. Recurrent networks and their LSTM variant consume sequences natively, convolutional networks scan for local patterns, and transformer-style models power large-scale work, but nearly all of that lives off-platform, where real training infrastructure exists. On-chart neural scripts stay tiny by necessity: a perceptron or a one-to-two hidden-layer MLP, weights either pre-trained offline and pasted in as constants, or nudged bar by bar in an online learning loop. Some indicators wearing the neural label are really fixed-weight formulas that never learn at all, which is worth checking before trusting the branding.

The training realities are unglamorous and decisive. Inputs need consistent scaling computed causally; class imbalance skews naive accuracy; and non-stationarity means a fit decays, motivating scheduled retraining under walk-forward rules. Raw outputs are scores, not probabilities, until checked by calibration, and averaging several independently initialized networks, an application of ensemble voting, is the cheap way to tame run-to-run variance. The Library spans the spectrum honestly: Noldo's ANN MACD ships offline-trained weights for BTC, while capissimo's perceptron strategy learns on the chart itself.

How to identify what a neural indicator actually does

Neural branding covers everything from real learned models to dressed-up moving averages. A short audit separates them.

1. 1List the inputs and their normalization: what features feed the network, and is the scaling computed from past data only or from the whole visible chart.
2. 2Count parameters against training examples: a few hundred weights learned from a few hundred effective examples is memorization wearing a lab coat.
3. 3Find where training happens: offline with weights embedded as constants, online with per-bar updates, or nowhere, because the weights are hand-set and nothing is learned.
4. 4Check the retraining and normalization for lookahead: full-history rescaling or per-bar refits that touch future data rewrite the past and inflate the backtest.
5. 5Ask how it was validated: walk-forward or out-of-sample metrics are evidence; a beautiful full-history equity curve on the training data is not.

How traders use it

• As a direction classifier: oscillator readings, returns, and volatility measures feed a small MLP whose output is read as the probability that price closes higher or lower over some horizon, then plotted as a signal line or entry arrows.
• As a learned composite: the network compresses several familiar inputs into one adaptive confluence score, which is then treated like any oscillator, with zones, crossovers, and divergence reads.
• As a filter on a base system: a rule-based setup only fires when the network's reading agrees, on the theory that the model captures conditions the fixed rules miss.
• With repaint discipline: models retrained every bar or normalized over the full visible history can rewrite past outputs, so repaint-safe engineering (frozen weights, causal normalization) is what keeps the historical track honest.
• As a sizing input rather than a trigger: a calibrated probability output scales position size or skips marginal setups, which uses the model's graded confidence instead of flattening it into arrows.

Neural Networks vs other learned models